arXiv subjects

Julien Velcin

Publications and source records attributed to Julien Velcin.

At least 19 recordsLinked to original sources

arXiv2026

Debias-SparseGPT: Bias-Aware Pruning for Large Language Models

Model compression techniques such as pruning and quantization facilitate the efficient deployment and acceleration of Large Language Models (LLMs). However, recent studies show that weight sparsification methods, such as SparseGPT, can amplify existing biases in models, with outputs varying significantly depending on persona cues in the prompt. In this paper, we introduce Debias-SparseGPT, a post-training pruning method incorporating representational debiasing using a second-order term defined over demographically contrasting inputs. We perform empirical validation of our method over a wide range of generative LLMs. Across models and sparsity regimes (25%, 50%, and structured 2:4 sparsity), Debias-SparseGPT consistently reduces pruning-induced bias compared to SparseGPT while preserving model perplexity and zero-shot accuracy. Under the most restrictive 2:4 structured sparsity pattern, which most aggressively degrades model quality, augmenting the calibration set with long-context, content-rich examples further improves both downstream performance and fairness. Overall, Debias-SparseGPT advances the bias-performance trade-off while preserving the computational efficiency of sparse models.

cs.CL↗

arXiv2026

The Fairness Collapse Phenomenon: Bias Amplification in Language Models Trained on Synthetic Data

Generative models trained on artificially generated data have been shown to exhibit model collapse, resulting in significant performance degradation. As synthetic content increasingly contaminates the training corpora of language models, this raises critical concerns about the use of open data in continued pretraining. Although previous work has demonstrated model collapse in language models, it remains unclear whether exposure to synthetic data amplifies or attenuates the social biases already present in pretrained models. Because language models are known to reproduce and amplify demographic stereotypes, recursive training on self-generated data may create a self-reinforcing feedback loop in which biased associations become progressively stronger across generations. We call this hypothesized phenomenon fairness collapse. In this work, we construct controlled training regimes in which models are repeatedly trained on synthetic data using the Bias in Bios dataset. Across experiments, we observe a consistent and concerning pattern: fairness degradation emerges before substantial degradation is reflected by standard language-modeling metrics. This result highlights a critical risk associated with synthetic data contamination in language model training: bias can increase silently before strong indicators of model collapse become apparent.

cs.CL↗

arXiv2026

Beyond Arrow's Impossibility: Fairness as an Emergent Property of Multi-Agent Collaboration

Fairness in language models is typically studied as a property of a single, centrally optimized model. As large language models become increasingly agentic, we propose that fairness emerges through interaction and exchange. We study this via a controlled hospital triage framework in which two agents negotiate over three structured debate rounds. One agent is aligned to a specific ethical framework via retrieval-augmented generation (RAG), while the other is either unaligned or adversarially prompted to favor demographic groups over clinical need. We find that alignment systematically shapes negotiation strategies and allocation patterns, and that neither agent's allocation is ethically adequate in isolation, yet their joint final allocation can satisfy fairness criteria that neither would have reached alone. Aligned agents partially moderate bias through contestation rather than override, acting as corrective patches that restore access for marginalized groups without fully converting a biased counterpart. We further observe that even explicitly aligned agents exhibit intrinsic biases toward certain frameworks, consistent with known left-leaning tendencies in LLMs. We connect these limits to Arrow's Impossibility Theorem: no aggregation mechanism can simultaneously satisfy all desiderata of collective rationality, and multi-agent deliberation navigates rather than resolves this constraint. Our results reposition fairness as an emergent, procedural property of decentralized agent interaction, and the system rather than the individual agent as the appropriate unit of evaluation.

cs.CL↗

arXiv2025

HatePrototypes: Interpretable and Transferable Representations for Implicit and Explicit Hate Speech Detection

Optimization of offensive content moderation models for different types of hateful messages is typically achieved through continued pre-training or fine-tuning on new hate speech benchmarks. However, existing benchmarks mainly address explicit hate toward protected groups and often overlook implicit or indirect hate, such as demeaning comparisons, calls for exclusion or violence, and subtle discriminatory language that still causes harm. While explicit hate can often be captured through surface features, implicit hate requires deeper, full-model semantic processing. In this work, we question the need for repeated fine-tuning and analyze the role of HatePrototypes, class-level vector representations derived from language models optimized for hate speech detection and safety moderation. We find that these prototypes, built from as few as 50 examples per class, enable cross-task transfer between explicit and implicit hate, with interchangeable prototypes across benchmarks. Moreover, we show that parameter-free early exiting with prototypes is effective for both hate types. We release the code, prototype resources, and evaluation scripts to support future research on efficient and transferable hate speech detection.

cs.CL↗

arXiv2025

Fair-GPTQ: Bias-Aware Quantization for Large Language Models

The high memory demands of generative language models have drawn attention to quantization, which reduces memory usage by mapping model weights to lower-precision integers. However, recent empirical studies show that, while efficient, quantization can increase the likelihood of generating biased outputs and degrade performance on fairness benchmarks. In this work, we draw new links between quantization and model fairness by adding explicit group-fairness constraints to the quantization objective and introduce Fair-GPTQ, the first quantization method explicitly designed to reduce unfairness in large language models. The added constraints guide the learning of the rounding operation toward less-biased text generation for protected groups. Specifically, we focus on stereotype generation involving occupational bias and discriminatory language spanning gender, race, and religion. Fair-GPTQ has minimal impact on performance, preserving at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to a half-precision model, and retains the memory and speed benefits of 4-bit quantization.

cs.CL↗

arXiv2025

Histoires Morales: A French Dataset for Assessing Moral Alignment

Aligning language models with human values is crucial, especially as they become more integrated into everyday life. While models are often adapted to user preferences, it is equally important to ensure they align with moral norms and behaviours in real-world social situations. Despite significant progress in languages like English and Chinese, French has seen little attention in this area, leaving a gap in understanding how LLMs handle moral reasoning in this language. To address this gap, we introduce Histoires Morales, a French dataset derived from Moral Stories, created through translation and subsequently refined with the assistance of native speakers to guarantee grammatical accuracy and adaptation to the French cultural context. We also rely on annotations of the moral values within the dataset to ensure their alignment with French norms. Histoires Morales covers a wide range of social situations, including differences in tipping practices, expressions of honesty in relationships, and responsibilities toward animals. To foster future research, we also conduct preliminary experiments on the alignment of multilingual models on French and English data and the robustness of the alignment. We find that while LLMs are generally aligned with human moral norms by default, they can be easily influenced with user-preference optimization for both moral and immoral data.

cs.CL↗

arXiv2024

Capturing Style in Author and Document Representation

A wide range of Deep Natural Language Processing (NLP) models integrates continuous and low dimensional representations of words and documents. Surprisingly, very few models study representation learning for authors. These representations can be used for many NLP tasks, such as author identification and classification, or in recommendation systems. A strong limitation of existing works is that they do not explicitly capture writing style, making them hardly applicable to literary data. We therefore propose a new architecture based on Variational Information Bottleneck (VIB) that learns embeddings for both authors and documents with a stylistic constraint. Our model fine-tunes a pre-trained document encoder. We stimulate the detection of writing style by adding predefined stylistic features making the representation axis interpretable with respect to writing style indicators. We evaluate our method on three datasets: a literary corpus extracted from the Gutenberg Project, the Blog Authorship Corpus and IMDb62, for which we show that it matches or outperforms strong/recent baselines in authorship attribution while capturing much more accurately the authors stylistic aspects.

cs.CL↗

arXiv2024

When Quantization Affects Confidence of Large Language Models?

Recent studies introduced effective compression techniques for Large Language Models (LLMs) via post-training quantization or low-bit weight representation. Although quantized weights offer storage efficiency and allow for faster inference, existing works have indicated that quantization might compromise performance and exacerbate biases in LLMs. This study investigates the confidence and calibration of quantized models, considering factors such as language model type and scale as contributors to quantization loss. Firstly, we reveal that quantization with GPTQ to 4-bit results in a decrease in confidence regarding true labels, with varying impacts observed among different language models. Secondly, we observe fluctuations in the impact on confidence across different scales. Finally, we propose an explanation for quantization loss based on confidence levels, indicating that quantization disproportionately affects samples where the full model exhibited low confidence levels in the first place.

cs.CL↗

arXiv2023

Mini Minds: Exploring Bebeshka and Zlata Baby Models

In this paper, we describe the University of Lyon 2 submission to the Strict-Small track of the BabyLM competition. The shared task is created with an emphasis on small-scale language modelling from scratch on limited-size data and human language acquisition. Dataset released for the Strict-Small track has 10M words, which is comparable to children's vocabulary size. We approach the task with an architecture search, minimizing masked language modelling loss on the data of the shared task. Having found an optimal configuration, we introduce two small-size language models (LMs) that were submitted for evaluation, a 4-layer encoder with 8 attention heads and a 6-layer decoder model with 12 heads which we term Bebeshka and Zlata, respectively. Despite being half the scale of the baseline LMs, our proposed models achieve comparable performance. We further explore the applicability of small-scale language models in tasks involving moral judgment, aligning their predictions with human values. These findings highlight the potential of compact LMs in addressing practical language understanding tasks.

cs.CL↗

arXiv2023

Dynamic Mixed Membership Stochastic Block Model for Weighted Labeled Networks

Most real-world networks evolve over time. Existing literature proposes models for dynamic networks that are either unlabeled or assumed to have a single membership structure. On the other hand, a new family of Mixed Membership Stochastic Block Models (MMSBM) allows to model static labeled networks under the assumption of mixed-membership clustering. In this work, we propose to extend this later class of models to infer dynamic labeled networks under a mixed membership assumption. Our approach takes the form of a temporal prior on the model's parameters. It relies on the single assumption that dynamics are not abrupt. We show that our method significantly differs from existing approaches, and allows to model more complex systems --dynamic labeled networks. We demonstrate the robustness of our method with several experiments on both synthetic and real-world datasets. A key interest of our approach is that it needs very few training data to yield good results. The performance gain under challenging conditions broadens the variety of possible applications of automated learning tools --as in social sciences, which comprise many fields where small datasets are a major obstacle to the introduction of machine learning methods.

cs.LG↗

arXiv2022

Multivariate Powered Dirichlet Hawkes Process

The publication time of a document carries a relevant information about its semantic content. The Dirichlet-Hawkes process has been proposed to jointly model textual information and publication dynamics. This approach has been used with success in several recent works, and extended to tackle specific challenging problems --typically for short texts or entangled publication dynamics. However, the prior in its current form does not allow for complex publication dynamics. In particular, inferred topics are independent from each other --a publication about finance is assumed to have no influence on publications about politics, for instance. In this work, we develop the Multivariate Powered Dirichlet-Hawkes Process (MPDHP), that alleviates this assumption. Publications about various topics can now influence each other. We detail and overcome the technical challenges that arise from considering interacting topics. We conduct a systematic evaluation of MPDHP on a range of synthetic datasets to define its application domain and limitations. Finally, we develop a use case of the MPDHP on Reddit data. At the end of this article, the interested reader will know how and when to use MPDHP, and when not to.

cs.LG↗

arXiv2022

Dirichlet-Survival Process: Scalable Inference of Topic-Dependent Diffusion Networks

Information spread on networks can be efficiently modeled by considering three features: documents' content, time of publication relative to other publications, and position of the spreader in the network. Most previous works model up to two of those jointly, or rely on heavily parametric approaches. Building on recent Dirichlet-Point processes literature, we introduce the Houston (Hidden Online User-Topic Network) model, that jointly considers all those features in a non-parametric unsupervised framework. It infers dynamic topic-dependent underlying diffusion networks in a continuous-time setting along with said topics. It is unsupervised; it considers an unlabeled stream of triplets shaped as \textit{(time of publication, information's content, spreading entity)} as input data. Online inference is conducted using a sequential Monte-Carlo algorithm that scales linearly with the size of the dataset. Our approach yields consequent improvements over existing baselines on both cluster recovery and subnetworks inference tasks.

cs.LG↗

arXiv2022

Explainable Clustering via Exemplars: Complexity and Efficient Approximation Algorithms

Explainable AI (XAI) is an important developing area but remains relatively understudied for clustering. We propose an explainable-by-design clustering approach that not only finds clusters but also exemplars to explain each cluster. The use of exemplars for understanding is supported by the exemplar-based school of concept definition in psychology. We show that finding a small set of exemplars to explain even a single cluster is computationally intractable; hence, the overall problem is challenging. We develop an approximation algorithm that provides provable performance guarantees with respect to clustering quality as well as the number of exemplars used. This basic algorithm explains all the instances in every cluster whilst another approximation algorithm uses a bounded number of exemplars to allow simpler explanations and provably covers a large fraction of all the instances. Experimental results show that our work is useful in domains involving difficult to understand deep embeddings of images and text.

cs.AI↗

arXiv2022

Serialized Interacting Mixed Membership Stochastic Block Model

Last years have seen a regain of interest for the use of stochastic block modeling (SBM) in recommender systems. These models are seen as a flexible alternative to tensor decomposition techniques that are able to handle labeled data. Recent works proposed to tackle discrete recommendation problems via SBMs by considering larger contexts as input data and by adding second order interactions between contexts' related elements. In this work, we show that these models are all special cases of a single global framework: the Serialized Interacting Mixed membership Stochastic Block Model (SIMSBM). It allows to model an arbitrarily large context as well as an arbitrarily high order of interactions. We demonstrate that SIMSBM generalizes several recent SBM-based baselines. Besides, we demonstrate that our formulation allows for an increased predictive power on six real-world datasets.

cs.LG↗

arXiv2022

Properties of Reddit News Topical Interactions

Most models of information diffusion online rely on the assumption that pieces of information spread independently from each other. However, several works pointed out the necessity of investigating the role of interactions in real-world processes, and highlighted possible difficulties in doing so: interactions are sparse and brief. As an answer, recent advances developed models to account for interactions in underlying publication dynamics. In this article, we propose to extend and apply one such model to determine whether interactions between news headlines on Reddit play a significant role in their underlying publication mechanisms. After conducting an in-depth case study on 100,000 news headline from 2019, we retrieve state-of-the-art conclusions about interactions and conclude that they play a minor role in this dataset.

cs.SI↗

arXiv2022

Interactions in Information Spread

Large quantities of data flow on the internet. When a user decides to help the spread of a piece of information (by retweeting, liking, posting content), most research works assumes she does so according to information's content, publication date, the user's position in the network, the platform used, etc. However, there is another aspect that has received little attention in the literature: the information interaction. The idea is that a user's choice is partly conditioned by the previous pieces of information she has been exposed to. In this document, we review the works done on interaction modeling and underline several aspects of interactions that complicate their study. Then, we present an approach seemingly fit to answer those challenges and detail a dedicated interaction model based on it. We show our approach fits the problem better than existing methods, and present leads for future works. Throughout the text, we show that taking interactions into account improves our comprehension of information interaction processes in real-world datasets, and argue that this aspect of information spread is should not be neglected when modeling spreading processes.

cs.SI↗

arXiv2022

Le Processus Powered Dirichlet-Hawkes comme A Priori Flexible pour Clustering Temporel de Textes

The textual content of a document and its publication date are intertwined. For example, the publication of a news article on a topic is influenced by previous publications on similar issues, according to underlying temporal dynamics. However, it can be challenging to retrieve meaningful information when textual information conveys little. Furthermore, the textual content of a document is not always correlated to its temporal dynamics. We develop a method to create clusters of textual documents according to both their content and publication time, the Powered Dirichlet-Hawkes process (PDHP). PDHP yields significantly better results than state-of-the-art models when temporal information or textual content is weakly informative. PDHP also alleviates the hypothesis that textual content and temporal dynamics are perfectly correlated. We demonstrate that PDHP generalizes previous work --such as DHP and UP. Finally, we illustrate a possible application using a real-world dataset from Reddit.

cs.CL↗

arXiv2021

Monitoring geometrical properties of word embeddings for detecting the emergence of new topics

Slow emerging topic detection is a task between event detection, where we aggregate behaviors of different words on short period of time, and language evolution, where we monitor their long term evolution. In this work, we tackle the problem of early detection of slowly emerging new topics. To this end, we gather evidence of weak signals at the word level. We propose to monitor the behavior of words representation in an embedding space and use one of its geometrical properties to characterize the emergence of topics. As evaluation is typically hard for this kind of task, we present a framework for quantitative evaluation. We show positive results that outperform state-of-the-art methods on two public datasets of press and scientific articles.

cs.CL↗